Why does ChatGPT forget my research notes?

ChatGPT forgets research notes because each chat opens in an isolated 128K-token context window with no automatic access to other chats, and Memory only stores short paraphrased notes around 8,000 tokens for your entire account. Citations, quotes, and reasoning chains get compressed to single sentences or dropped. The fix is to keep research notes in a persistent project memory layer.

Research is high-density text: quotes, page numbers, citations, counter-arguments. ChatGPT is not built to store that.

1. Chats are isolated context windows. A new chat opens with a 128K window on GPT-4o (smaller on older tiers), and nothing from yesterday's chat is pre-loaded. The research you reasoned through is on OpenAI's servers as raw text, but the model is not given any of it on a fresh load.

2. Memory does not store research, it sketches it. ChatGPT's Memory feature writes short notes about you ("user is researching battery chemistries"). It cannot store the actual passages, page numbers, or quotes you found. By the time you ask "what was the figure on cycle life from the Chen 2024 paper", Memory has only the vague fact that you were reading about batteries.

3. Uploaded sources expire. Any PDFs you uploaded to support a chat sit in that chat's sandbox and get recycled after roughly 3 hours of idle time. A new chat has no PDF to consult, even if the previous chat was full of citations from it.

The net effect: the chat sidebar shows weeks of work, the model knows almost none of it.

Research compounds when memory holds, and falls apart when it does not:

• Citations evaporate. "We were going to cite Chen 2024 on figure 3" becomes a sentence ChatGPT no longer has. You either re-find the paper or you fudge the cite.
• Counter-arguments reset. The careful list of objections you walked through last week is gone. ChatGPT cheerfully suggests the option you already eliminated.
• Long-term lit reviews fragment. A literature review that spans months should sharpen; without memory it forks into parallel half-versions across different chats.

The fix is not "keep one giant chat open". Long chats slow down, hit length caps, and self-trim. The fix is to detach research memory from chat memory.

OpenAI ships three partial solutions for sustained research.

Memory stores short notes across the account. Useful for "user studies battery chemistry". Useless for "the cycle-life figure in Chen 2024 was 1,200 cycles at 80% retention". Notes are paraphrases; you cannot quote from them.

Projects (Pro / Plus / Team) group chats, files, and instructions into a research folder. A real improvement for cross-chat continuity inside one product. Still capped at a per-project file limit, still subject to the same Memory store, still locked inside ChatGPT.

Reference chat history lets ChatGPT search recent chats and pull snippets. It is best-effort, region-limited on some plans, and silently misses older or differently-worded mentions.

OpenAI's Memory FAQ is the canonical reference on what Memory does and does not preserve. None of these features are designed to hold a literature review.

For a single class paper, the natives are enough. For a real research workflow, they are not.

Research rarely lives in one tool. You read in Perplexity, draft in ChatGPT, argue in Claude, and cite in a writing app. Each tool builds its own thin memory of your project and none of them share it. The lit review you grew in ChatGPT is invisible to Claude, and the rebuttals you sharpened in Claude are invisible to ChatGPT.

The way out is a unified research store that any AI you use can read from on the fly.

MemoryLake turns each research thread into a Project with proper memory types — Fact Memory for citations, Event Memory for the research timeline, Conversation Memory for the reasoning you logged.

• Quotes and citations as Fact Memory. Page numbers, quotes, figures, and source URLs are stored verbatim, not paraphrased. Conflict detection flags contradictions across sources instead of silently averaging them.
• MemoryLake-D1 parses real PDFs. Tables, multi-column layouts, and scanned figures from journal articles are extracted with visual plus logical verification, so a chart caption is not silently dropped.
• Plug into 4,000+ open research datasets. Built-in access to PubMed, arXiv, SEC EDGAR, USPTO patents, and clinical trials means ChatGPT can pull from primary sources without you copy-pasting them in.

MemoryLake holds the top published LoCoMo long-context score of 94.03%, retrieves at millisecond latency, and runs AES-256 end-to-end encryption so even MemoryLake cannot read your sources.

1. Create a project and load your research. Sign in to MemoryLake, open Project Management, click Create Project, and name it for the thread ("Lit review — solid-state batteries"). Upload PDFs, notes, and reference exports to the Document Drive. Add key citations and findings as named entries in the Memories tab so they ride along with the project.
2. Generate an MCP Server endpoint. Open the MCP Servers tab in the project, click Add MCP Server, name it "ChatGPT integration", and click Generate. MemoryLake returns an API key ID, secret, and endpoint URL. Copy the secret immediately — it is shown only once.
3. Connect ChatGPT. Browser ChatGPT does not yet support MCP, so call the REST API with your Bearer token to pull relevant sources and prior reasoning into each chat, or paste a system prompt that points ChatGPT to your MemoryLake project ID. The Python SDK can also log new chat findings back as Conversation Memory automatically.